/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x4-minmax-avx.c | 62 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local 115 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local 155 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local
|
D | up16x4-minmax-fma3.c | 62 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local 115 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local 155 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local
|
D | up16x4-minmax-fma3-acc2.c | 62 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local 118 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local 160 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local
|
D | up16x4-minmax-avx-acc2.c | 62 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local 118 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local 160 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local
|
D | up8x4-minmax-fma3.c | 62 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3() local 103 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3() local
|
D | up8x4-minmax-avx.c | 62 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx() local 103 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx() local
|
D | up8x4-minmax-avx-acc2.c | 62 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx_acc2() local 105 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx_acc2() local
|
D | up8x4-minmax-fma3-acc2.c | 62 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3_acc2() local 105 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3_acc2() local
|
D | up16x9-minmax-fma3.c | 87 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local 185 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local 255 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
|
D | up16x9-minmax-avx.c | 87 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local 185 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local 255 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
|
D | up16x9-minmax-fma3-acc2.c | 87 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 188 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 260 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
|
D | up8x9-minmax-avx.c | 87 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx() local 158 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx() local
|
D | up8x9-minmax-fma3.c | 87 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3() local 158 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3() local
|
D | up8x9-minmax-avx-acc2.c | 87 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx_acc2() local 160 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx_acc2() local
|
D | up8x9-minmax-fma3-acc2.c | 87 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3_acc2() local 160 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3_acc2() local
|
D | up8x25-minmax-avx-acc2.c | 167 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local 336 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
|
D | up8x25-minmax-fma3-acc2.c | 167 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local 336 __m256 vacc01234567p0 = _mm256_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
|
/external/XNNPACK/src/f16-dwconv/gen/ |
D | up16x4-minmax-neonfp16arith.c | 62 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local 104 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local 130 float16x8_t vacc01234567p0 = vld1q_f16(w); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local
|
D | up16x4-minmax-neonfp16arith-acc2.c | 62 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local 107 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local 135 float16x8_t vacc01234567p0 = vld1q_f16(w); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local
|
D | up8x4-minmax-neonfp16arith.c | 62 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith() local 88 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith() local
|
D | up8x4-minmax-neonfp16arith-acc2.c | 62 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2() local 90 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2() local
|
D | up16x9-minmax-neonfp16arith.c | 87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local 164 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local 210 float16x8_t vacc01234567p0 = vld1q_f16(w); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
|
D | up16x9-minmax-neonfp16arith-acc2.c | 87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local 167 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local 215 float16x8_t vacc01234567p0 = vld1q_f16(w); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
|
D | up8x9-minmax-neonfp16arith.c | 87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith() local 133 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith() local
|
D | up8x9-minmax-neonfp16arith-acc2.c | 87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2() local 135 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2() local
|